A Feasibility Study of Reinforcement Learning for Adaptive Urban Traffic Signal Control

Urban traffic congestion remains a major challenge for modern cities, leading to increased travel delays, fuel consumption, and harmful vehicular emissions. Conventional traffic signal control strategies, such as fixed-time, actuated, and delay-based systems, rely on predefined logic and local thresholds, which limits their ability to adapt to dynamic and non-linear traffic conditions. As urban networks grow in complexity and sustainability becomes a critical objective, more adaptive and intelligent control approaches are required. This thesis investigates the application of reinforcement learning (RL) for adaptive traffic signal control, with a focus on jointly improving traffic efficiency and reducing emissions. Three RL approaches Q-Learning, Deep Q-Networks (DQN), and Proximal Policy Optimisation (PPO) are developed and evaluated using a high-fidelity microscopic simulation framework based on the Simulation of Urban Mobility (SUMO) platform with HBEFA-based emission modelling. Both single-intersection and multi-intersection corridor scenarios are examined under moderate and high traffic demand conditions derived from realistic urban traffic data. The traffic signal control problem is formulated as a Markov Decision Process, where signal controllers act as learning agents that observe local traffic states, including queue lengths, lane occupancy, and signal phases, and select control actions to optimize a multi-objective reward function. Unlike many prior studies, this work explicitly incorporates fuel consumption and CO2 emissions into the learning objective, encouraging smoother traffic flow and environmentally efficient operation. For multi-intersection networks, an independent multi-agent framework is employed, allowing intersections to coordinate implicitly through shared traffic dynamics. The results demonstrate that RL-based controllers consistently outperform traditional signal control methods across all evaluated scenarios. DQN achieves the strongest overall performance, delivering substantial reductions in delay, queue length, fuel consumption, and emissions, particularly under high-demand conditions. PPO provides stable and robust improvements, while tabular Q-Learning yields moderate gains but is limited in larger state spaces. Importantly, the findings confirm that improvements in traffic efficiency directly translate into environmental benefits, indicating that mobility and sustainability objectives can be jointly optimised. Overall, this thesis provides strong empirical evidence that reinforcement learning offers a scalable and effective solution for adaptive traffic signal control, supporting its potential application in future intelligent and environmentally sustainable urban traffic management systems.